Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Template-Guided 3D Molecular Pose Generation via Flow Matching and Differentiable Optimization
Authors: Noémie Bergues, Arthur Carré, Paul Join-Lambert, Brice Hoffmann, Arnaud Blondel, Hamza Tajmouati
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present a two-stage method for ligand conformation generation guided by such templates. In the first stage, we introduce a molecular alignment approach based on flow-matching to generate 3D coordinates for the ligand, using the template structure as a reference. In the second stage, a differentiable pose optimization procedure refines this conformation based on shape and pharmacophore similarities, internal energy, and, optionally, the protein binding pocket. We introduce a new benchmark of ligand pairs co-crystallized with the same target to evaluate our approach and show that it outperforms standard docking tools and open-access alignment methods, especially in cases involving low similarity to the template or high ligand flexibility. |
| Researcher Affiliation | Collaboration | 1 Iktos SA, 65 rue de Prony, 75017 Paris, France 2 Institut Pasteur, Université Paris Cité, UMR 3528, Paris, France 3 Doctoral school MTCI (ED 563), Paris Cité University, Paris, France |
| Pseudocode | Yes | Algorithm S1: INFERENCE Input: Query 2D graph Gquery, template molecule Mtemplate, number of sampling steps N, number of samples K Algorithm S2: TRAINING Input: Query molecules with 2D graphs [G1 query, ..., GK query], true coordinates [C1, ..., CK] and associated template molecules [M1 template, ..., MK template], number of epochs Nepochs, learning rate α. Algorithm S3: HARMONICSAMPLING Input: 2D Molecular graph G Algorithm S4: FMA Input: Molecular graph G with adajency matrix X, nodes features fnode, edges features fedge, time t, template-query positions x RNatoms 3 at time t, N MHA blocks , N MHA heads , N V F N blocks , N V F N heads . Algorithm S5: VECTORFIELDNETWORK Input: Scalar features fs RNnodes cn, vector features fv RNnodes 3 cn, edges features fe RNedges ce, query-template positions x, time t, Nblocks, Nheads Algorithm S6: GATEDUPDATE Input: Scalar features fs RNnodes cn, vector features fv RNnodes 3 cn, output dimension d Algorithm S7: POSEOPTIMIZATION Input: Ligand conformation with coordinates x, number of optimization steps niterations, learning rate lr |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: No, we do not provide open access to the code, as it is proprietary and developed within a commercial context. However, we will release the Align Dock Bench benchmark to support reproducibility and comparative evaluation. |
| Open Datasets | Yes | To evaluate the performance of our method, we introduce Align Dock Bench, a new benchmark comprising 369 protein-ligand (PL) template query pairs. ... The benchmark is available at Zenodo 2. |
| Dataset Splits | Yes | Align Dock Bench. The benchmark includes 369 PL query structures, each associated with a corresponding PL template structure. Training set. ... The resulting training set contains 301, 348 complex pairs covering 111, 678 unique molecules. To prevent train test leakage, we excluded any training molecule with a Morgan fingerprint [Rogers and Hahn, 2010] Tanimoto similarity greater than 0.5 to any ligand in Align Dock Bench (Figure S7). |
| Hardware Specification | Yes | We trained FMA for 100 epochs with a batch size of 128 on a single NVIDIA GeForce RTX 2080 Ti GPU. 4Experiments were run on a single NVIDIA T4 Tensor Core GPU. |
| Software Dependencies | Yes | Docking baselines include Vina (v1.2.5) [Trott and Olson, 2010] and r Dock (v24.04.204-legacy) [Ruiz-Carmona et al., 2014], while alignment baselines include Fit Dock (v1.0.9) [Yang et al., 2022], LS-align (Version J201704171741) [Hu et al., 2018], and ROSHAMBO [Atwi et al., 2024]. |
| Experiment Setup | Yes | We trained FMA for 100 epochs with a batch size of 128 on a single NVIDIA GeForce RTX 2080 Ti GPU. We used the Adam W optimizer with a learning rate of 3 10 4 and a weight decay of 10 5. Hyperparameters are detailed in Table S1. Optimization Objective. The optimization objective combines all scores into a weighted loss function defined as follows: Loptim = α SSTS(Mquery, Mtemplate) β SPTS(Mquery, Mtemplate) ω Spocket(Mquery, Ptemplate) + γ Einternal(Mquery). Further details on the optimization algorithm and hyperparameters are provided in Appendix B. |